Resources
Methods, records and ground rules, kept in the open
A result is only as good as the method behind it and the record that shows what was done. This page says where the lab keeps both, how we share what we know, and the rules we follow when AI helps with the work.
Protocols
A new lab member should not have to learn a method by guessing what a methods section left out.
I wrote step-by-step protocols from the methods of my own papers. They cover cell culture and neuronal differentiation, chicken embryo work, organoids, imaging, electrophysiology, western blots, reporter assays, small RNA, drug measurement in tissue, and statistics. Each one separates what the paper reported from general laboratory practice, quotes catalogue numbers as printed, and ends with a list of what the paper did not say.
The methods are published and were used in the work behind my papers. The step-by-step versions here are version 0.1, written out from those published methods sections, so they carry any gap a methods section leaves. Methods that collaborators performed are marked as reference methods. A protocol moves to version 1.0 when someone has run it from the page, written down what they changed, and a second person has repeated it. Validated versions will be deposited on protocols.io, so that papers can cite them by DOI.
How to use them
- Read the whole protocol before you touch anything, including the section on approvals.
- Run a pilot. Write down every change and the reason for it.
- Ask for the approvals your own project needs. The permits named in the papers belong to the original studies.
- Found an error or a gap? Open an issue or a pull request. Corrections are welcome from anyone.
Text is licensed CC BY 4.0. Reuse and adapt it with credit.
Records: SciSure (eLabNext)
Protocols say what to do. The notebook says what was done.
Lab work at the University of Oslo is recorded in SciSure for Research, the university's electronic lab notebook. It was formerly called eLabJournal and is also known as eLabNext. Everyone in the lab writes in it, so the record is shared, searchable and kept in one place.
The GitHub protocols are the master text. A notebook entry names the protocol and the version it used, so any result can be traced back to its method.
What every entry holds
- The question the experiment asks, and what result would change the plan.
- The result expected and the point at which the experiment will stop, both written before it starts.
- Anything that surprised you, even if it is not yet explained.
- The protocol and version, with every deviation from it.
- Materials, with lot numbers for the ones that matter.
- Where the raw data are stored.
- The result, and the next step.
A surprise is only visible against an expectation recorded beforehand. A result that matches your expectation may only mean you stopped as soon as it did.
The test for a good entry: someone else can repeat the work from the page alone.
Using AI in research
AI tools can help with parts of the work. They cannot take responsibility for it.
These are the rules in this lab. They follow the European Commission's guidelines on generative AI in research, the European Code of Conduct for Research Integrity, and University of Oslo guidance. Funders and journals differ in the details, and the rules are changing, so check the call or the journal before you submit. Last reviewed November 2026.
Do
- Take responsibility for every claim, figure, reference and line of code, whoever or whatever drafted it.
- Open every reference you cite and check that it says what you claim. Look up each one in PubMed, Crossref or the publisher's page, check it has not been retracted, and add citations by hand.
- Tell the funder, the journal and your co-authors when AI did more than fix the language, in the words their policy asks for. Put it in the Methods or acknowledgements.
- Keep a record of substantial AI use in your SciSure entry, using the AI-use log in the lab handbook. Keep AI-assisted code under version control with tests.
- Use only university-approved tools for unpublished, personal or patient-related material, matched to the data's protection class.
- Use AI for language editing and translation without embarrassment, especially in a second language. Check afterwards that it has not dropped quotation marks, citations or numbers.
- Practise reading, analysis, writing and troubleshooting unaided, and bring AI in after you have made a first attempt. You should be able to explain and redo any AI-assisted analysis or code without the tool.
- Before you trust AI-written analysis code on new data, run it on a dataset you have already analysed by hand and get the same answer.
- Confirm an AI-derived result with a second method or an experiment before it goes into a figure.
- Keep the question, the experimental design, the bench work and the sign-off with people.
Do not
- List an AI tool as an author.
- Use AI to make up or change data, results, references, or images of experimental data.
- Use generative AI to create images for a manuscript. A labelled schematic for a talk or the website is fine if it is not presented as data.
- Use AI to enhance an image (denoising, deblurring, super-resolution) without recording the model, version, parameters and seed. Treat enhancement as analysis, and say so in the methods.
- Upload unpublished work, personal data or anyone else's confidential material to an external tool.
- Put a manuscript or a grant proposal you are reviewing into any AI tool, or let AI write or judge a review.
- Run an AI note-taker or transcription tool in a lab meeting or call about unpublished or confidential work unless everyone present agrees.
- Present AI output as checked fact: a summary of a paper, a statistical result, or code that "ran fine".
- Hide substantial use of AI.
- Let AI replace your own reading, reasoning and bench work.
- Take an AI's verdict on novelty or importance in place of an expert's.
- Use an AI-text detector as evidence against a student or colleague. These tools are inaccurate and biased, and they can flag a person's own text after machine translation.
- Give an AI agent credentials, write access to shared data or code, or control of an instrument. Agents that act on their own need the PI's approval first, with a defined scope and a log. Nothing runs at the bench or on lab equipment without a named person approving each protocol.
What a human scientist brings
Choosing the question is the first contribution, and I keep it with people because the evidence that AI chooses good questions is thin. In one study of language-model research, ideas from an AI model were rated more novel than experts' ideas before anyone ran them. After experts carried them out, the AI ideas lost more ground than the experts' did, and the two sets ended up statistically indistinguishable.1 A small biology study points the same way: in one simulated molecular-genetics task, run twice, ChatGPT-4 proposed fewer hypotheses than human subjects and never noticed an anomaly.5 That is one task and one model.
Creativity matters, and it is one piece of this. A new idea still has to be judged, tested and signed by someone. In the independent test cited below, some ideas the system marked as novel were already well known.3
Hands and judgement come next. Wet-lab work still needs a person at the bench, who sees that the cells look wrong before any number says so. In a published case of an AI research agent, people wrote the lab protocols and ran the bench experiments, and the authors say the agent does not yet produce protocols precise enough to execute.2
The same paper has a useful warning about references. When the literature-search agents were removed, the base model's proposals contained many invented references, which a human reference check found.2 Check every reference yourself.
AI systems also misreport their own work. An independent test of an automated "AI scientist" found experiments that failed on coding errors, manuscripts with invented numbers, and ideas marked novel that were already in the literature.3 The developers' own published account lists hallucinated citations and incorrect implementations among its failure modes, says only one of its three submitted papers was accepted, at a workshop where the bar is lower than at the main conference, and says people filtered the outputs before submission. Its own supplementary example papers contain claims that the authors' tables contradict.4 These tests used machine-learning tasks and models from 2024 and 2025, not wet-lab biology, so read them as failure modes to look for, not as current error rates.
Last comes accountability. A signature, a retraction notice and a conversation with a patient group all need a name behind them. Every policy cited here says the same: the human stays responsible.
Where AI helps
- Language editing and translation for people writing in a second language.
- Finding and sorting literature, when you then read what it found. Search tools lean towards well-indexed journals and languages, so check against PubMed or another database.
- Image analysis and structure prediction, where the tools are validated and the method is reported.
- Drafting and explaining code, when you can read the code and a test shows it works.
- Analysis with machine learning, when you start from a plain baseline, split the data so that no subject, sample or site appears on both sides, and report the uncertainty.
I have not claimed numbers here because they date quickly. The sources below are the place to look.
Sources and where to check the rules
- European Commission, Living guidelines on the responsible use of generative AI in research
- ALLEA, European Code of Conduct for Research Integrity (2023), doi:10.26356/ECOC
- European Research Council, position on AI
- University of Oslo, guidelines for the use of AI
- ICMJE, use of AI by authors and COPE, authorship and AI tools, doi:10.24318/cCVRZBms
- Blau W. et al., Protecting scientific integrity in an age of generative AI. Proc. Natl Acad. Sci. USA 121, e2407886121 (2024). doi:10.1073/pnas.2407886121
- Rolauffs B. et al., Rethinking the scientific method in the age of AI. Nat. Rev. Bioeng. 4, 2–3 (2026). doi:10.1038/s44222-025-00386-3
- 1. Si C., Hashimoto T. & Yang D., The Ideation-Execution Gap: execution outcomes of LLM-generated versus human research ideas. ICLR 2026; preprint arXiv:2506.20803. doi:10.48550/arXiv.2506.20803
- 2. Ghareeb A. E. et al., A multi-agent system for automating scientific discovery. Nature 655, 497–505 (2026). doi:10.1038/s41586-026-10652-y
- 3. Beel J., Kan M.-Y. & Baumgart M., Evaluating Sakana's AI Scientist: bold claims, mixed results, and a promising future? ACM SIGIR Forum 59(1), 1–20 (2025). doi:10.1145/3769733.3769747; preprint arXiv:2502.14297
- 4. Lu C. et al., Towards end-to-end automation of AI research. Nature 651, 914–919 (2026). doi:10.1038/s41586-026-10265-5
- 5. Ding A. W. & Li S., Generative AI lacks the human creativity to achieve scientific discovery from scratch. Sci. Rep. 15, 9587 (2025). doi:10.1038/s41598-025-93794-9